fix(stt): correct off-by-one in MultiSpeakerAdapter RMS timerange window

livekit-agents/livekit/agents/stt/multi_speaker_adapter.py

@@ -270,7 +270,7 @@ def _get_rms_for_timerange(self, start_time: float, end_time: float) -> float |
 
         start = int((self._pushed_duration - start_time) / self._frame_size)
         end = int((self._pushed_duration - end_time) / self._frame_size)
-        start = len(self._rms_buffer) - start - 1
+        start = len(self._rms_buffer) - start
         end = len(self._rms_buffer) - end
 
         if end < 0 or start >= len(self._rms_buffer):

tests/test_multi_speaker_rms_timerange.py

@@ -0,0 +1,81 @@
+import pytest
+
+from livekit import rtc
+from livekit.agents.stt.multi_speaker_adapter import (
+    PrimarySpeakerDetectionOptions,
+    _PrimarySpeakerDetector,
+)
+
+pytestmark = pytest.mark.unit
+
+
+def _constant_frame(value: int, *, sample_rate: int, samples_per_channel: int) -> rtc.AudioFrame:
+    """Build a mono int16 frame where every sample equals ``value``."""
+    sample = value.to_bytes(2, "little", signed=True)
+    return rtc.AudioFrame(
+        data=sample * samples_per_channel,
+        sample_rate=sample_rate,
+        num_channels=1,
+        samples_per_channel=samples_per_channel,
+    )
+
+
+class TestGetRmsForTimerange:
+    def test_excludes_frame_before_start_time(self) -> None:
+        """The RMS window for [start_time, end_time) must not include the frame
+        immediately preceding start_time.
+        """
+        opts = PrimarySpeakerDetectionOptions(frame_size_ms=100, min_rms_samples=1)
+        detector = _PrimarySpeakerDetector(primary_detection_options=opts)
+
+        sample_rate = 1000
+        samples_per_frame = 100  # 100ms at 1000Hz, matches frame_size_ms above
+
+        # frame 1: silence, covers [0.0, 0.1)
+        detector.push_audio(
+            _constant_frame(0, sample_rate=sample_rate, samples_per_channel=samples_per_frame)
+        )
+        # frame 2: loud, covers [0.1, 0.2)
+        detector.push_audio(
+            _constant_frame(1000, sample_rate=sample_rate, samples_per_channel=samples_per_frame)
+        )
+
+        # query only the loud frame's range; the silent frame must not leak in
+        rms = detector._get_rms_for_timerange(0.1, 0.2)
+
+        assert rms == pytest.approx(1000.0)
+
+    def test_window_contains_exactly_the_requested_frame_count(self) -> None:
+        """A [start_time, end_time) window spanning N frames must return exactly
+        N samples, not N+1 (which would happen if the frame before start_time
+        leaked into the window).
+        """
+        opts = PrimarySpeakerDetectionOptions(frame_size_ms=100, min_rms_samples=1)
+        detector = _PrimarySpeakerDetector(primary_detection_options=opts)
+
+        sample_rate = 1000
+        samples_per_frame = 100  # 100ms at 1000Hz
+
+        for _ in range(5):  # silence
+            detector.push_audio(
+                _constant_frame(0, sample_rate=sample_rate, samples_per_channel=samples_per_frame)
+            )
+        for _ in range(5):  # loud
+            detector.push_audio(
+                _constant_frame(
+                    1000, sample_rate=sample_rate, samples_per_channel=samples_per_frame
+                )
+            )
+
+        # query the last 5 frames pushed (the loud ones), anchored on the detector's
+        # own clock so this isn't sensitive to float drift in accumulated durations
+        end_time = detector._pushed_duration
+        start_time = end_time - 5 * detector._frame_size
+
+        # exactly 5 frames fall inside the window; requiring 5 samples should succeed...
+        detector._opt.min_rms_samples = 5
+        assert detector._get_rms_for_timerange(start_time, end_time) == pytest.approx(1000.0)
+
+        # ...but requiring 6 should fail, since only 5 frames actually fall in range.
+        detector._opt.min_rms_samples = 6
+        assert detector._get_rms_for_timerange(start_time, end_time) is None